Abstract: Working based on preset rules often only achieves local optimization with a single goal, resulting in low efficiency in active distribution network reconstruction. To address the above issues, a dynamic reconstruction of the distributed energy high penetration scenario active distribution network based on deep reinforcement learning is proposed. Clarify the abstract framework of the problem, define the state, action space, optimization objectives, and constraints around this framework, and then construct a dynamic reconstruction problem model for the distribution network. Using deep Q-networks, construct graph structures and perform graph convolution operations, while combining loss functions to achieve mapping between states and actions based on deep reinforcement learning. Divide the distribution network into multiple control areas and deploy independent agents to achieve active dynamic reconstruction of the distribution network through information exchange and action coordination. The experiment shows that the research method achieves an effective balance of multiple objectives in the dynamic reconstruction of active distribution networks. Compared with the comparative method, it has better convergence speed and final convergence value, which improves the efficiency of active distribution network reconstruction.